v0.3.3

0.3.3

Patch Changes

• chore(renovate): drop dead @kernhq automerge rule

v0.3.2

0.3.2

Patch Changes

• chore: check the packed tarball can resolve its own imports

v0.3.1

0.3.1

Patch Changes

• be1ffc9: Scope the example seeder's guard to the workspace, and import eq.

  The guard left workspace_id to row-level security, which is exactly the mistake the template
  should not be teaching: the transaction is bound to the workspace, so RLS scopes it on a
  correctly-configured instance and silently does not on one whose owner can bypass a policy. Every
  first-party seeder had the same defect and every workspace after the first was created empty.

v0.3.0

0.3.0

Minor Changes

• 2832ea0: Show how a module fills a workspace created with example content.

  demo: { seed } is the extension point; the kernel subscribes it for you. The example is deliberately
  two rows and a guard, because the guard is the part a module author has to get right: delivery is
  at-least-once, so a seeder can be asked twice for one workspace, and the check is "is this workspace
  empty?" rather than a marker somebody has to maintain.

v0.2.17

0.2.17

Patch Changes

• build: force one @kernhq/kernel copy with pnpm.overrides

v0.2.16

0.2.16

Patch Changes

• 89feda5: Resolve a single @kernhq/kernel copy

  The template pulled two kernels into a consumer's tree: @kernhq/testing@0.1.12 declared a
  @kernhq/kernel: ^0.9.0 it never imported, and @kernhq/ui@0.14.0 depends on ^0.9.0 too. With
  the module's own kernel at ^0.10.3, both floors are raised so everything resolves one copy —
  @kernhq/testing to ^0.1.14, which dropped the phantom dependency, and @kernhq/ui to ^0.14.4,
  which depends on ^0.10.2.

  Two structurally identical declarations of ServerModule are not assignable to each other, so a
  duplicate kernel is a type error in whatever hosts the module — and it is invisible in the umbrella
  workspace, which pins one copy regardless.

v0.2.15

0.2.15

Patch Changes

• b15e900: Require @kernhq/kernel ^0.10.3, whose rlsPolicySql emits the policy drop

  rlsPolicySql omitted the drop policy if exists before its create policy until
  @kernhq/kernel@0.10.3, so an author following the documented path got a policy that could not be
  applied twice. The kernel is fixed; this raises the floor so a fresh checkout of the template
  resolves a kernel where the helper and migrations/0001_rls.sql actually agree.

  The floor is what makes it true rather than the range alone: degit copies pnpm-lock.yaml, and
  pnpm does not re-resolve a version that still satisfies its range — so ^0.10.0 kept installing
  0.10.0. The migration comment and README step 5 now say the helper emits all four statements, dated
  to the kernel version that made it so.

v0.2.14

0.2.14

Patch Changes

• af8faeb: Make 0001_rls.sql replayable, and ship the migration test that proves it

  create policy has no if not exists, so applying the migration folder a second time answered
  policy "notes_ws_isolation" for table "notes" already exists. A module's migrations are the first
  thing the kernel runs, so that is a host service that never binds its port — not a broken feature.
  The policy is now preceded by drop policy if exists, as it is in every first-party module.

  src/server/migrations.test.ts is new and ships in the tarball. It applies the folder to a database
  created from nothing, applies it again, and asks the Postgres catalogue which tables carrying
  workspace_id are actually secured — so a tenant table added without a policy fails by name rather
  than being silently readable. It needs a DATABASE_URL it can create databases from; CI already
  starts one.

v0.2.13

0.2.13

Patch Changes

• fe17fa3: The client module documents overlays, the contribution a navigation does not destroy.

  Everything else the template declares — the route, the sidebar, the widget — lives inside a page,
  so opening something else unmounts it. @kernhq/kernel@0.10.2 adds overlays for the cases where
  that is wrong: something in progress, an upload, a countdown. The example is commented out, because
  a module with nothing to keep should ship no overlay rather than an empty one, and it names the two
  rules that are easy to get wrong — an overlay is mounted on every page of the workspace, so it must
  draw nothing until it has something to say, and it is handed no location, because the route it was
  mounted on is not the route the person is on now.

v0.2.12

0.2.12

Patch Changes

• 1e80aaa: Peer @kernhq/contracts@^0.8.0, which adds archivedAt to WorkspaceSummary. A caret on 0.x does
  not cross a minor, so the previous ^0.7.0 could not reach it. The template is what a third party
  starts a module from, so a stale peer here is the first thing they would hit.